A policy tuned for one robot often behaves differently on another, whether due to the sim-to-real gap, unknown payloads, or the differing dynamics of two instances of the same robot. In contact-rich, dynamic manipulation, even small motion discrepancies can result in failure to track reference motion, since they disrupt the timing and modes of contact. Common remedies, such as domain randomization or system identification, either produce overly conservative task policies or require data that must be recollected for each robot or payload. We introduce the Torque Adaptation Module (TAM), a learned module that adapts the torque commands sent to the robot to match the behavior of an ideal robot. TAM operates between the low-level controller that tracks the policy's actions and the robot's torque interface. It includes a history encoder that embeds proprioceptive history into a latent state and a torque adaptor that computes residual torque corrections. Because TAM depends only on proprioceptive history and not on policy observations, or the action space, the same TAM weights can be reused to adapt policies with different action spaces (joint targets, end-effector targets, or direct torques). The policies themselves do not need to be trained with domain randomization of robot parameters. Instead, we offload the need for domain randomization to TAM by training it entirely in randomized simulation, using multi-robot pretraining followed by a robot-specific fine-tuning step that still requires no real-robot data. We evaluate TAM zero-shot on a real Franka Panda robot across dynamic manipulation tasks that include a vision-based box pushing policy (from RL), a flip policy (from BC), and an MPC ball-on-plate balancing. Our experiments show that TAM improves zero-shot real-robot execution compared to online system identification and RMA baselines and enables robust dynamic manipulation performance.

Robot manipulation in unstructured environments requires efficient and reliable Swept Volume Collision Detection (SVCD) for safe motion planning. Traditional discrete methods potentially miss collisions between these points, whereas SVCD continuously checks for collisions along the entire trajectory. Existing SVCD methods typically face a trade-off between efficiency and accuracy, limiting practical use. In this paper, we introduce NeuralSVCD, a novel neural encoder-decoder architecture tailored to overcome this trade-off. Our approach leverages shape locality and temporal locality through distributed geometric representations and temporal optimization. This enhances computational efficiency without sacrificing accuracy. Comprehensive experiments show that NeuralSVCD consistently outperforms existing state-of-the-art SVCD methods in terms of both collision detection accuracy and computational efficiency, demonstrating its robust applicability across diverse robotic manipulation scenarios. Code and videos are available at https://neuralsvcd.github.io/.

The usage of 3D vision algorithms, such as shape reconstruction, remains limited because they require inputs to be at a fixed canonical rotation. Recently, a simple equivariant network, Vector Neuron (VN) has been proposed that can be easily used with the state-of-the-art 3D neural network (NN) architectures. However, its performance is limited because it is designed to use only three-dimensional features, which is insufficient to capture the details present in 3D data. In this paper, we introduce an equivariant feature representation for mapping a 3D point to a high-dimensional feature space. Our feature can discern multiple frequencies present in 3D data, which, as shown by Tancik et al. (2020), is the key to designing an expressive feature for 3D vision tasks. Our representation can be used as an input to VNs, and the results demonstrate that with our feature representation, VN captures more details, overcoming the limitation raised in its original paper.

We present DEF-oriCORN, a framework for language-directed manipulation tasks. By leveraging a novel object-based scene representation and diffusion-model-based state estimation algorithm, our framework enables efficient and robust manipulation planning in response to verbal commands, even in tightly packed environments with sparse camera views without any demonstrations. Unlike traditional representations, our representation affords efficient collision checking and language grounding. Compared to state-of-the-art baselines, our framework achieves superior estimation and motion planning performance from sparse RGB images and zero-shot generalizes to real-world scenarios with diverse materials, including transparent and reflective objects, despite being trained exclusively in simulation. Our code for data generation, training, inference, and pre-trained weights are publicly available at: https://sites.google.com/view/def-oricorn/home.

A robot operating in a partially observable environment must perform sensing actions to achieve a goal, such as clearing the objects in front of a shelf to better localize a target object at the back, and estimate its shape for grasping. A POMDP is a principled framework for enabling robots to perform such information-gathering actions. Unfortunately, while robot manipulation domains involve high-dimensional and continuous observation and action spaces, most POMDP solvers are limited to discrete spaces. Recently, POMCPOW has been proposed for continuous POMDPs, which handles continuity using sampling and progressive widening. However, for robot manipulation problems involving camera observations and multiple objects, POMCPOW is too slow to be practical. We take inspiration from the recent work in learning to guide task and motion planning to propose a framework that learns to guide POMCPOW from past planning experience. Our method uses preference learning that utilizes both success and failure trajectories, where the preference label is given by the results of the tree search. We demonstrate the efficacy of our framework in several continuous partially observable robotics domains, including real-world manipulation, where our framework explicitly reasons about the uncertainty in off-the-shelf segmentation and pose estimation algorithms.

Our goal is to develop an efficient contact detection algorithm for large-scale GPU-based simulation of non-convex objects. Current GPU-based simulators such as IsaacGym and Brax must trade-off speed with fidelity, generality, or both when simulating non-convex objects. Their main issue lies in contact detection (CD): existing CD algorithms, such as Gilbert-Johnson-Keerthi (GJK), must trade off their computational speed with accuracy which becomes expensive as the number of collisions among non-convex objects increases. We propose a data-driven approach for CD, whose accuracy depends only on the quality and quantity of offline dataset rather than online computation time. Unlike GJK, our method inherently has a uniform computational flow, which facilitates efficient GPU usage based on advanced compilers such as XLA (Accelerated Linear Algebra). Further, we offer a data-efficient solution by learning the patterns of colliding local crop object shapes, rather than global object shapes which are harder to learn. We demonstrate our approach improves the efficiency of existing CD methods by a factor of 5-10 for non-convex objects with comparable accuracy. Using the previous work on contact resolution for a neural-network-based contact detector, we integrate our CD algorithm into the open-source GPU-based simulator, Brax, and show that we can improve the efficiency over IsaacGym and generality over standard Brax. We highly recommend the videos of our simulator included in the supplementary materials.

We propose a novel fast and accurate simulation framework for contact-intensive tight-tolerance robotic assembly tasks. The key components of our framework are as follows: 1) data-driven contact point clustering with a certain variable-input network, which is explicitly trained for simulation accuracy (with real experimental data) and able to accommodate complex/non-convex object shapes; 2) contact force solving, which precisely/robustly enforces physics of contact (i.e., no penetration, Coulomb friction, maximum energy dissipation) with contact mechanics of contact nodes augmented with that of their object; 3) contact detection with a neural network, which is parallelized for each contact point, thus, can be computed very quickly even for complex shape objects with no exhaust pair-wise test; and 4) time integration with PMI (passive mid-point integration), whose discrete-time passivity improves overall simulation accuracy, stability, and speed. We then implement our proposed framework for two widely-encountered/benchmarked contact-intensive tight-tolerance tasks, namely, peg-in-hole assembly and bolt-nut assembly, and validate its speed and accuracy against real experimental data. It is worthwhile to mention that our proposed simulation framework is applicable to other general contact-intensive tight-tolerance robotic assembly tasks as well. We also compare its performance with other physics engines and manifest its robustness via haptic rendering of virtual bolting task.

Although, in the task of grasping via a data-driven method, closed-loop feedback and predicting 6 degrees of freedom (DoF) grasp rather than conventionally used 4DoF top-down grasp are demonstrated to improve performance individually, few systems have both. Moreover, the sequential property of that task is hardly dealt with, while the approaching motion necessarily generates a series of observations. Therefore, this paper synthesizes three approaches and suggests a closed-loop framework that can predict the 6DoF grasp in a heavily cluttered environment from continuously received vision observations. This can be realized by formulating the grasping problem as Hidden Markov Model and applying a particle filter to infer grasp. Additionally, we introduce a novel lightweight Convolutional Neural Network (CNN) model that evaluates and initializes grasp samples in real-time, making the particle filter process possible. The experiments, which are conducted on a real robot with a heavily cluttered environment, show that our framework not only quantitatively improves the grasping success rate significantly compared to the baseline algorithms, but also qualitatively reacts to a dynamic change in the environment and cleans up the table.

The object manipulation is a crucial ability for a service robot, but it is hard to solve with reinforcement learning due to some reasons such as sample efficiency. In this paper, to tackle this object manipulation, we propose a novel framework, AP-NPQL (Non-Parametric Q Learning with Action Primitives), that can efficiently solve the object manipulation with visual input and sparse reward, by utilizing a non-parametric policy for reinforcement learning and appropriate behavior prior for the object manipulation. We evaluate the efficiency and the performance of the proposed AP-NPQL for four object manipulation tasks on simulation (pushing plate, stacking box, flipping cup, and picking and placing plate), and it turns out that our AP-NPQL outperforms the state-of-the-art algorithms based on parametric policy and behavior prior in terms of learning time and task success rate. We also successfully transfer and validate the learned policy of the plate pick-and-place task to the real robot in a sim-to-real manner.

In this paper, we propose a novel sim-to-real framework to solve bolting tasks with tight tolerance and complex contact geometry which are hard to be modeled. The sim-to-real has desirable features in terms of cost and safety, however, that of the assembly task is rare due to the lack of simulator, which can robustly render multi-contact assembly. We implement the sim-to-real transfer of nut tightening policy which is adaptive to uncertain bolt positions. This can be realized through developing a novel contact model, which is fast and robust to complex assembly geometry, and novel hierarchical controller with reinforcement learning (RL), which can perform the tasks with a narrow and complicated path. The fast and robust contact model is achieved by utilizing configuration space abstraction and passive midpoint integrator (PMI), which render the simulator robust even in a high stiffness contact condition. And we use sampling-based motion planning to construct a path library and design linear quadratic tracking controller as a low-level controller to be compliant and avoid local optima. Additionally, we use the RL agent as a high-level controller to make it possible to adapt to the bolt position uncertainty, thereby realizing sim-to-real. Experiments are performed to verify our proposed sim-to-real framework.

We propose a novel data-driven learning-based contact clustering (i.e., of contact points and contact normals) framework for rigid-body robot simulation, with its accuracy established/verified by real experimental data. We first construct an experimental robotic setup with force/torque (F/T) sensors to collect real contact motion/force data. We then design a multilayer perceptron (MLP) network for the contact clustering based on the full motion and force/torque information of the contacts. We also adopt the constraint-based optimization contact solver to facilitate the learning of our MLP network during the training. Our proposed data-driven/learning-based contact clustering framework is then verified against the experimental setup, compared with other techniques/simulators and shown to significantly (or meaningfully) enhance the accuracy of contact simulation as compared to them.

Electrical motor and hydraulic actuation widely-used in robotics are “internal actuation” with their actuators sitting at the joint between two links. This internal actuation is fundamentally limiting to construct a large-size dexterously-articulated robot, since any external force (and its own link weight) is to be accumulated to the base multiplied by the moment arm length, requiring extremely strong/sturdy base actuator/structure as the system size increases. In this paper, we propose a novel robotic system, LASDRA (large-size aerial skeleton with distributed rotor actuation), which, by utilizing distributed rotors as “external actuation”, can overcome this limitation of internal actuation and enables us to realize large-size dexterously-articulated robots. We present its design and modeling, joint locking strategy to increase its loading capability, and also a novel decentralized control scheme to allow for compliant operation with scalability against the number of links. Trajectory tracking and valve turning experiments are also performed to validate the theory.